Fault Neural Classifier Applied to a Level Control Real System

Transcription

1 Fault Neural Classifier Applied to a Level Control Real System Raphaela G. Fernandes, Diego R. C. Silva Universidade Federal do Rio Grande do Norte Natal-RN, Brazil raphaela@dca.ufrn.br, diego@dca.ufrn.br Luiz A. H. G. de Oliveira, Adrião D. D. Neto Universidade Federal do Rio Grande do Norte Natal-RN, Brazil affonso@dca.ufrn.br, adriao@dca.ufrn.br Abstract This work presents the development of a fault detection and isolation system applied to a system of real levels connected by Fieldbus network. The system was developed using artificial neural networks and tested just as much in a simulated environment as in real environment intending to highlight difficulties found in real tests which there is not when one works only with simulation. I. INTRODUCTION With the arrival of the industrial automation, each time more grows the necessity of guaranteeing itself security and trustworthiness of the equipment used in the execution of the processes. In a dynamic system, the faults can cause alterations in parameters (or variable) critical of the system, or even changes in its dynamics. This way, the faults can be detected, isolated and eliminated, providing more confidence for the system. In other words, a system that has a tolerant behavior to faults must be capable of automatically detecting and classifying faults occurred, just as it is capable to continue functioning afterwards. The faults that can come to occur in a dynamic system or in an industrial plant can generate damages so much that can harm the equipment belonging to the procedure, putting at risk the workers involved in it. The necessity to avoid these types of damages increases the importance of having an efficient fault detection and isolation system (FDI). There are many methods already consolidated of detection and isolation of faults. Some of them are based on the physical redundancy, in other words, in the duplication of hardware components, like sensors, actuators and controllers [1]. In these cases, the outlets of the equivalent pairs of components are the initial point for a possible abnormal detection. The subtraction of these two values is called, in the technical jargon, residue, which in turn, is one of the basic pairs in the FDI systems [1]. Basically, if the outputs values of the pairs are close, or, the residue shows to be very close to zero, it means that there is not a fault. There is a divergence between them and also if a considerable residue then, there is a fault. The principle disadvantage of these methods is the use of extra equipment that can increase the maintenance at a high rate, besides this it needs more physical space for its accommodation. In the seventies arose other techniques based on the paradigm of analytic redundancy. With this procedure, one can use the system model to generate redundancy signals which are compared with signals measured from the physical sensors [2]. Similarly to the physical redundancy the outputs of the sensors are compared to the ones calculated through the mathematical model with the object of generating the residues, and consequentially, if there is a significant difference between them, it characterizes the occurrence of a fault. One of the difficulties with this procedure is the necessity of having a mathematical model very close to reality. In recent years, researches made in the context of Faults Detection and Isolation system (FDI) presented procedures that uses the principles of Computer Intelligence, such as, Fuzzy Logic system [3] and Artificial Neural Networks (ANNs) [4]. This article describes the project and implementation of a fault detection and isolation system applied to a dynamic system in real time, for instance, a plant composed of two tanks (plant of levels), in which one of the tanks has its level controlled. Basically, the FDI system was divided in two parts: the first corresponds to neural identification of the plant model and the second, to the detection and isolation ( or classification) of faults in process. In this paper, we first turned the system parameters for a simulated environment, then we used there parameters to detect and to isolate faults in a real system. It is showed that the FDI system was able to accuse the faults at the moment they are ocurring, surpassing several difficulties found only in real environment, such as: noises and unmodeling dynamic. II. METHODOLOGIE DEVELOPMENT In this paper a fault detection and isolation system using artificial neural networks was developed and applied to a level system. It was tested using the mathematical model of the system, available in the manufacturer s manual [9], and also the real physical plant. In this article will be shown only the results applied to the real plant. The general scheme of the functioning system is shown in Figure 1. In this case, while the level system is in execution, an elaborated system from the ANNs tries to find its identification using its inputs (x(k)). Each time, the output level system (y(k)) is compared to the output identification system (y (k)), generate a residue value (r(k) = y(k) y (k)) that later will be used in the fault isolation/classification system. Then one

2 analyzes the residue values and indicates the occurrence or not of faults and, in case that it occurs, indicates what type it is. Fig. 1. General scheme of FDI system Two ways for neural levels identification system were developed and tested [5], therefore, only one of them was applied to the faults detection and classification system. In the next section, the levels system will be presented and, afterwards, the following steps to the conclusion of the system FDI applied to same. III. USED PROCESS: LEVEL SYSTEM Figure 2 shows the level plant composed of two tanks in cascade, representing a second order model with an input. A Proportional Integrative Derivative (PID) [6] control strategy is applied in this plant to control the level of tank 2. The real data extracted from the training neural network plant had also been captured in supervisory system using the OPC (OLE for process control) [8] standard communication. The OPC is responsible for providing information about the industrial plant on a supervisory level for any computer of the network. Besides this, it allows the alterations of some parameters of the FF system configuration. A neural identification was made in two environments: simulated and real. To simulate the dynamic of the plant, a non-linear mathematical model was used as described in the equations 1 and 2 [9]. L 1 = a 1 A 1 2gL1 + K m A 1 V p (1) L 2 = a 2 2gL2 + a1 2gL1 (2) A 2 A2 Table I presents the descriptions and values of the parameters shown in equations 1 and 2. TABLE I PLANT PARAMETERS Name Description Value K m Pump Constant 4, 6(cm 3 /s)/v V p Tension applied to pump 22 < V p < 22 a 1 Tank 1 output diameter 0, cm a 2 Tank 2 output diameter 0, cm A 1 Tank 1 area cm 2 A 2 Tank 2 area cm 2 g Gravity acelerator 980cm/s 2 Finally, the level plant model together with the controller PID is illustrated in Figure 3. As it is possible to see, the reference signal is the level desired for tank 2. The desired signal (r(k)) is compared to the actual level from the tank 2 (L2(k)), generating an error (e(k)) which is used to the PID controller to generate a control signal (V p ) to the pump that injects water to tank 1. Fig. 2. Arquitecture from the real environment. The water that flows out from the small hole of tank 1 falls into tank 2. This in turn, also has a small hole through which the water falls directly to the reservoir. A pump is responsible for taking the water from the reservoir to the tank 1. In each tank there is a Foundation Fieldbus (FF) pressure sensor [7], used to measure the corresponding levels connected to the Fieldbus network. Besides the pressure sensors, a FF/ loop of current from 4 to 20 ma converter is used to inject signals to the water pump. The industrial network Foundation Fieldbus is connected to a supervisory computer through Ethernet network interfaces. From this computer is carried out all the devices configuration process and later supervised. Fig. 3. Level plant with PID controller In the following section, it will be presented the developing of the neural networks which will make identification, detection and isolation of faults in the level system. IV. LEVEL SYSTEM NEURAL IDENTIFICATION The neural identification system was defined as identification in two steps, which means the existence of an ANN to evaluate the level of tank 1 and another to evaluate the level of tank 2. The training of the ANNs can be illustrated by figures 4 and 5. The architecture of the ANN 1, used to evaluate the level of tank 1, was:

3 Fig. 4. First Stage: Determination of tank 1 level from the system input the fault isolation/classification stage, if the output of ANN 1 diverges from the real output of tank 1, for example, it means that a fault could not have occurred on the side and sensors from the process that evolves tank 2, and vice-versa, that is, reducing the possibilities at the moment of classifying the fault. With identification in two steps, it is possible to get two residues, R 1 and R 2, where R 1 = L 1 L 1 and R 2 = L 2 L 2. The fault isolation strategy can be seen in Figure 6. In this case, a ANN, denominated ANN 3, is trained receiving as input data the values from R 1 and R 2. The networks output corresponds to a vector of N numbers, where N is a quantity of faults that the network will be able to classify. For example, considering N = 3, the possible outputs for the network shown on table II would be: Fig. 5. Second Stage: Determination of tank 2 level from the system input Three nodes in the input layer, where the inputs are: V p (k), V p (k 1) and L 1 (k 1); A hidden layer with 3 neurons with the sigmoid activation function; A neuron on the output layer (linear activation function), representing the level of tank 1 evaluating in the actual instant k, L 1(k). The ANN 2 architecture, used to evaluate the level of tank 2, was: Two nodes in the input layer which are L 1 (k 1) and L 2 (k 1) ; A hidden layer with 3 neurons with the sigmoid activation function. One neuron in the output layer (linear activation function ), representing the level of tank 2 evaluating in the actual instant k, L 2(k). To implement the strategy of neural level system identification, it used a Multiple Layers Perceptron (MLP) architecture [10]. The number of neurons from the hidden layers was chosen from initial tests and the neurons are completely connected. In the next section it will be presented the FDI system development in the level plants, which is based on the neural identification in the two steps showed before. V. FAULT DETECTION AND ISOLATION In accordance with the identification neural project considered, it was verified therefore that it would be useful for Fig. 6. General project for fault detection and isolation TABLE II POSSIBLE OUTPUTS FOR THE CLASSIFICATION OF 3 FAULTS. Normal Situation Fault Fault Fault Considering the two residues to detect and isolate the faults, the FDI system could detect in maximum 8 different faults, in which the situation R 1 = 0 and R 2 = 0 would be a normal behavior of the system. With the intention of testing, we managed to foresee only six types of faults with joint distinct residues. Table III shows the five types of faults selected for the result analysis, where + represents the positive residue, the negative residue and 0 the equal residue or very close to zero. TABLE III DISPOSITION OF THE 6 FAULTS RESIDUES Fault Description R 1 R 2 Absence of fault New hole in tank No fall of water in tank 2 2 Decrease of hole in tank Decrease of hole in tank Increase of hole in tank Extra flow of water in tank 1 + 0

4 As the neural classifier works based on positive values, negative and null values for the residues, it could have been translated to a set of rules with conditional sentences. For example, if R 1 was positive and R 2 negative, it would be sufficient to make a conditional test to indicate that fault 2 is occurring. On the other hand, if another dynamic system could be modeled for N residues, where N is a very high value, it would be a little laborious to codify several conditional judgments to inform the fault. Besides this, the neural can classify faults with similar residues ( two different faults with the same R 1 and R 2, for example, R 1 = + and R 2 = 0), however with different amplitudes, without the necessity of a new conditional test verifying the threshold of the fault. VI. NEURAL IDENTIFICATION RESULTS The data captured for training had been generated from the normal functioning of the plant levels, that is, it applied reference signals for the level of tank 2 and the control signal was generated from PID controller. The reference signals were always unit degree or sinusoidal type, with amplitude varying through the time. Figure 7 shows one of the data sets used for training. In the first chart, there is the value of tank 1 throughout the time. The second chart shows the level of tank 2 in blue and the reference used by the PID controller in red. The third chart shows that the controller signal is being applied to the pump, generated by the controller. Fig. 8. Example - Identification in two steps - real environment Based on the results that are presented in Figure 8, it can be inferred that, for both situations, the errors between the level of the tanks and the generated ones in the identification are bigger during the transitory period of the tank 2 level. This is due to the result of a lot of noises being present during the test and the pump has components with non-linear dynamic, making difficult the learning of neural networks. The three following diagrams showed in Figures 9, 10 and 11 present other results from the neural identification of the system. They allow an analysis of what is closest to the real system function, if the mathematical model or the ANN is trained from the real data. It would be worth while to point out that the gains to the PID controller were applied as much to the real plant controller as to the mathematical model. The first chart of Figure 9 presents the level of tank 1 of the real system (in blue), the same generated by ANN 1 with it is identification (in red) and the level generated by the mathematical model considering the same reference. The second chart shows the errors between the real level and the one generated by the mathematical model (in blue), and the real level and the one generated by the identification ANN 1 (in red). Fig. 7. Example of data set for training - real environment After the training of the two networks used for the neural identification, the medium square errors found for the ANN 1 and the ANN 2 were approximately, 10 4 and 10 5, respectively. An example of neural identification for data not presented in the training of the networks can be seen in Figure 8. On the first chart are the levels of tank 1 real (blue) and the one generated by ANN 1 (red) throughout the time. On the second chart it is easy to see the level of tank 2 real (blue), generated by ANN 2 (red) and the reference to be followed by the level of this same tank (green). Finally, in the last chart are presented the residue values R 1 (blue) and R 2 (red). Fig. 9. The level differences of tank 1: mathematical model, real and generated by ANN

5 Figure 10, the first chart shows the level of tank 2 in the real system (in blue), the same generated by ANN 2 with the identification (in red) and the level generated by the mathematical model considering the same reference. The same way, the second diagram shows the errors between the real level and the one generated by mathematical model (in blue), and the real level that is generated by the identification ANN 2 (in red). VII. FAULT DETECTION AND ISOLATION RESULTS The neural network, ANN 3, used to detect and isolate faults was trained only by using the neural identification results using the mathematical model of system [5]. With the mathematical model, faults were simulated and the residues captured, allowing the data generation suggesting the normal function of the system or the occurrence of fault. This approach has been chosen due to the fact that in real situations, where one does not have a data base with the information of occurring faults all the time, it could be complicated to generate faults in the system in order to be able to develop an application that could detect them. Figure 12 shows a output of the supervisor program developed to monitor the system operation. It is possible to observe the occurrence or the non-occurrence of the system s faults according to the strategy described in section V. Fig. 10. The level differences of tank 2: mathematical model, real and generated by ANN Finally, Figure 11 exhibits the signs of the control generated by the real function of the plant (first chart, in blue) and by the mathematical model (first chart, in red), as well as the difference between them (second chart). Fig. 12. Supervisory Program Fig. 11. The control signal differences: mathematical model and real From these last three charts, it can be concluded that the data generated by the neural identification are much closer to reality than that generated by the mathematical model available in the level of plant manual. It can also be seen that with the mathematical model the level of tank 2 stabilizes more quickly, due to the fact that the mathematical equations modeled an ideal situation, without considering noises and a highly nonlinearity of the pump (that in fact happens in the real case). In the next section results from the detection and isolation system will be presented as well as the fault level system isolation. In real time, instant values are exhibited from the real tanks level, as well as the estimated levels by the neural networks and the residue values. The supervisor program also shows a graph with the evaluation of residue values throughout time. In the case that faults occur, a corresponding alarm with the fault will be highlighted in red and a message informing the fault is displayed. The three graphs of Figures 13, 14 and 15 illustrate the moment when the faults of types 2, 3 and 5 occur (see table II). At this moment, the residues diverge the values close to zero and the ANN 3 is able to identify and classify the fault. At this point the fault corresponding button changes to color red and the messages "Water increase in tank 1 and water decrease to tank 2", "Water increase in tank 2 and water decrease in tank 1" and "Water increase in tank 2" are shown as faults 2, 3 and 6, respectively. As already mentioned, the strategy of the FDI system was implemented as much in the simulated environment as was in the real one. Many difficulties have been discovered while working in the real system, as for example: tests should have been made considering that the water in tank 1 would

6 not have overflowed: the necessity of using extra equipment (power supply, module amplifier, etc.); the presence of leaks between hoses and pressure measures; problems with the neural identification at transitory moments due to the nonlinearity of the pump (mathematically modeled to an ideal form), among others. Fig. 13. Fault 2 - Decrease of small hole in Tank 1 VIII. FINAL CONSIDERATIONS AND EXPECTATIONS This paper showed the development of a fault detection and isolation system. It was applied to a real system which was connected through an industrial network. This system was divided into two parts: neural identification and classification/isolation of the faults. It is executed in the course of real time and is able to accuse the faults the moment that the abnormalities occur, surpassing all difficulties not found in the simulated environment. Real system identification results with the mathematical model and with neural networks were discussed, showing that the neural results showed were closer to reality. This is due to the fact that the mathematical model presents an ideal situation of the system functioning, what does not occur in reality. Perspectively, it intends to improve the neural identification for situations in which the PID controller still has not established the level of tank 2 (transitory). Besides this, improve the neural classifier so that it can be able to detect and isolate faults that have similar behaviors to the residues R 1 and R 2. REFERENCES Fig. 14. Fault 3 - Decrease of small hole in Tank 2 [1] L. Weber, "A script for explanation of the basic concepts of tolerance of faults", [2] S. Persin, B. Tovornik, N. Muskinja and D. Valh, "Increasing process safety using analytical redundancy," Electrotechnical Review, 2002, pp [3] F.J. Uppal, R.J. Patton and M. Witczak, "A neuro-fuzzy multiple-model observer approach to robust fault diagnosis based on the DAMADICS benchmark problem," Control Engineering Practice, [4] H.J. Shin, D. Eom and S. Kim "One-class support vector machines - an application in machine fault detection and classification," Computer and Industrial Engineering, 2005, pp [5] R. Fernandes, D. Silva, L.A. Guedes, and A. Neto, "Neural Identification of the Systems of Levels in the Foundation Fieldbus environment," Latin American Congress of automatics, Salvador-BA-Brazil, [6] Katsuhiko Ogata, Modern Control Engineering, LTC, [7] F. Lima, L. Guedes, A. Ortiz and A. Maitelli "Hybrid Enviroment for Tests and Training in Fieldbuses," VI Induscon, [8] Li Zheng and H. Nakagawa, "OPC (OLE for process control) specification and its developments," in 41st SICE Annual Conference, vol. II, 2002, pp [9] Coupled Water Tank Experiments. Quanser Innovate Educate. [10] Simon Haykin, Neural Networks: Principles and Practice, Bookman, Fig. 15. Fault 5 - Extra water low in Tank 1